Grok is xAI's assistant, reachable inside X, at grok.com, and through mobile apps. Type a request in plain language and refine it with follow-ups, the same as any chatbot. The one thing it does that the others do not is read public posts on X as they appear, which makes it the right tool for live sentiment and breaking events, and the wrong tool for anything where a Twitter-shaped sample would mislead you.
Getting in
Grok appears in the X sidebar if you have an account, and separately at grok.com and in the iOS and Android apps. The same conversation history follows you between them.
There is a limited free tier. Paid access runs through X Premium or through SuperGrok, xAI's standalone subscription, which as of August 2026 is priced at $30 a month or $300 a year. A much heavier tier, SuperGrok Heavy, sits at $300 a month and exists for people running long reasoning jobs and parallel agents rather than for writing social posts. Message allowances change often enough that the number in any guide is stale within weeks, so check your own account before planning around a limit.
The modes are the part worth learning
Most people type into the default box and never touch the controls, which is where most of the capability sits.
- Think. Switches Grok onto its reasoning path: it works through the problem internally before answering, and you can expand the thoughts to see the route it took. Slower, and clearly better on anything with multiple steps or a constraint that has to hold across the whole answer. The expanded reasoning is also the fastest way to spot where it went wrong.
- DeepSearch. Research mode. It searches, shows the steps, and cites what it used. Treat the citations as the output rather than the summary above them, because the summary is where an assistant is most likely to smooth over a source that did not quite say that.
- Voice. Speech to speech, useful for thinking out loud on a walk, poor for anything you need to keep.
The current flagship model is Grok 4.5, released in July 2026, with earlier versions still reachable through the API. As with every model family, the version number matters far less to your results than whether you gave it enough context.
Paste it into Grok's custom instructions and the brainstorm becomes a system: topic clusters, a calendar, briefs, and a voice that holds across a quarter rather than a session.
View Lena →What Grok is genuinely better at
One thing, and it is a real one: what people are saying right now. Because it reads public X posts as they are posted, it can tell you how a launch is landing, what the reaction to an announcement looks like an hour in, or whether a complaint about your product is one person or a pattern. No amount of web search substitutes for that, because the conversation is not indexed as pages.
Practical uses that follow from it:
- Live sentiment on a brand, a competitor or an announcement.
- Catching up on a fast-moving story without reading four hundred posts.
- Finding the specific complaint people keep repeating, in their own words, which is the best copy research there is.
- Spotting whether a trend is real or three accounts amplifying each other.
Beyond that it is a competent general assistant: drafting, summarising, brainstorming, explaining code. Competent, not distinctive.
The blind spot that comes with the strength
Public posts on X are not a representative sample of anything except public posts on X. They skew toward the people with the strongest opinions and the time to post them, and the loudest cluster is often the smallest one.
So: use Grok to find out what is being said, and do not let it tell you what people think. "Sentiment is 70 percent negative" from a sample of self-selected posters is a sentence that has ended careers. When it hands you a read on public opinion, ask it how many distinct accounts that read rests on. The answer is often surprisingly small.
Prompts behave the same here
The structure that works everywhere works here: role, task, context, format. Grok's looser default tone means the format instruction matters slightly more, since without one it drifts chatty.
Two things worth adding specifically for Grok:
- When you want live data, say so explicitly. "Based on posts from the last 24 hours" changes the answer; without it you may get training-data recall dressed as current.
- Ask for the receipts. "Quote three actual posts" turns an unfalsifiable summary into something you can check in ten seconds.
Our free Prompt Optimizer rewrites a rough prompt into that structure, and it works in Grok, ChatGPT, Gemini and Claude. See also how to write better AI prompts.
Making images with Grok
Grok generates images too: ask in the same chat box, in plain language. It is quick and noticeably more permissive than its rivals, which cuts both ways — fewer refusals, and more output you would not put in front of a client. Treat it as a sketchpad for concepts rather than a production tool, be specific about style and composition, and keep text out of the image itself, because rendered text is still where every image model embarrasses itself. On paid tiers, Grok Imagine extends the same idea to short animated clips.
Grok, ChatGPT or Claude
The honest answer is that for most professional work the differences between the frontier models are smaller than the difference a good prompt makes, so pick on the thing that is not interchangeable.
- Grok when the question is about right now, or about what people on X are saying.
- Claude for long-form writing and work where the tone has to hold over thousands of words.
- ChatGPT for breadth of tooling and integrations.
The useful discipline is keeping your prompts portable so switching costs you nothing. See Claude vs Gemini for work if most of your day is writing and analysis. If the bulk of it is research with citations, here is how to use Perplexity for research.
A workflow that survives contact with real work
Start from a saved prompt rather than typing instructions from scratch. Add the specifics for today. Read the first answer sceptically, then ask for exactly one improvement at a time, because a request to fix four things at once usually fixes two and quietly breaks a third. When the answer is finally right, save that prompt, not the answer.
That last habit is the whole difference between people who find AI useful and people who find it inconsistent.
The saved-prompt habit above, done properly: one file carrying the role, the method and the output rules, so the answer is the same standard on a Monday as on a Friday.
View Lena →Tiers, pricing and model versions checked August 2026. xAI ships changes frequently; confirm current terms at x.ai before planning around them.
Want more? Browse the full prompt library. Every pack works in Grok, ChatGPT, Gemini and Claude.
